Nemotron Content Safety Audio Dataset

## Overview The Nemotron Content Safety Audio Dataset is a multimodal extension of Nvidia's Nemotron Content Safety Dataset V2 (Aegis 2.0). It contains 1,928 English-language audio files generated from the test-set prompts of Aegis 2.0, paired with their original text and safety labels across 23 violation categories. It is distributed as CSV metadata plus audio files and is intended for evaluating multimodal AI safety classifiers and audio guardrails. ## Schema - audio_filepath — string — relative path to the spoken-prompt audio file - text — string — original English text prompt - label — string — safe / unsafe classification - violation_categories — string/list — one or more of the 23 Aegis safety taxonomy categories - prompt_id — string — links back to source Aegis 2.0 record - duration — float — audio length in seconds - speaker / voice metadata — string — TTS voice identifier (where provided) - +additional metadata columns from the parent Aegis 2.0 record ## Sources - nvidia/Nemotron-Content-Safety-Audio-Dataset on Hugging Face — https://huggingface.co/datasets/nvidia/Nemotron-Content-Safety-Audio-Dataset — License: CC-BY-4.0 - Parent dataset: nvidia/Aegis-AI-Content-Safety-Dataset-2.0 - Reference paper: arXiv:2501.09004 ## Methodology Nvidia constructed the dataset by taking the test split of the Aegis 2.0 content-safety prompts (expert- and human-annotated adversarial and benign prompts spanning 23 violation categories) and synthesizing spoken-audio versions of each prompt via text-to-speech. The resulting audio files are aligned 1:1 with the source text records, preserving original safety labels and taxonomy assignments so that audio-modality safety classifiers can be evaluated against the same ground truth as the text-modality benchmark. ## Known gaps & limitations All prompts are English-only, so the dataset does not support multilingual audio safety evaluation. Audio is TTS-synthesized rather than recorded from natural speakers, so it under-represents accent, prosody, background noise, and disfluency seen in real-world audio attacks. The 1,928 file size is suitable for evaluation but small for training. Distribution of violation categories inherits any class imbalance present in the Aegis 2.0 test set. As a safety dataset it contains adversarial and harmful textual content by design. ## Intended use & out-of-scope - IS for: evaluating audio and multimodal guardrail/classifier models, red-teaming speech-to-text safety pipelines, benchmarking content-moderation systems on spoken inputs. - NOT for: training general-purpose ASR models, multilingual safety evaluation, or as a stand-alone training corpus for safety classifiers without the parent Aegis 2.0 train split. _PII signals: cc_shape×10 (Luhn-valid: 0) present in the sample. Common in public datasets (papers, logs) but worth knowing before joining with private data._ Original supplier listing: Nemotron Content Safety Audio Dataset (Aegis 2.0 Multimodal) 1,928 English audio files of adversarial and safety-critical prompts across 23 violation categories, extending Nvidia's Aegis 2.0 content-safety benchmark into the audio modality for multimodal guardrail evaluation.